]> CyberLeo.Net >> Repos - FreeBSD/FreeBSD.git/blob - sys/kern/kern_tc.c
This commit was generated by cvs2svn to compensate for changes in r156283,
[FreeBSD/FreeBSD.git] / sys / kern / kern_tc.c
1 /*-
2  * ----------------------------------------------------------------------------
3  * "THE BEER-WARE LICENSE" (Revision 42):
4  * <phk@FreeBSD.ORG> wrote this file.  As long as you retain this notice you
5  * can do whatever you want with this stuff. If we meet some day, and you think
6  * this stuff is worth it, you can buy me a beer in return.   Poul-Henning Kamp
7  * ----------------------------------------------------------------------------
8  */
9
10 #include <sys/cdefs.h>
11 __FBSDID("$FreeBSD$");
12
13 #include "opt_ntp.h"
14
15 #include <sys/param.h>
16 #include <sys/kernel.h>
17 #include <sys/sysctl.h>
18 #include <sys/syslog.h>
19 #include <sys/systm.h>
20 #include <sys/timepps.h>
21 #include <sys/timetc.h>
22 #include <sys/timex.h>
23
24 /*
25  * A large step happens on boot.  This constant detects such steps.
26  * It is relatively small so that ntp_update_second gets called enough
27  * in the typical 'missed a couple of seconds' case, but doesn't loop
28  * forever when the time step is large.
29  */
30 #define LARGE_STEP      200
31
32 /*
33  * Implement a dummy timecounter which we can use until we get a real one
34  * in the air.  This allows the console and other early stuff to use
35  * time services.
36  */
37
38 static u_int
39 dummy_get_timecount(struct timecounter *tc)
40 {
41         static u_int now;
42
43         return (++now);
44 }
45
46 static struct timecounter dummy_timecounter = {
47         dummy_get_timecount, 0, ~0u, 1000000, "dummy", -1000000
48 };
49
50 struct timehands {
51         /* These fields must be initialized by the driver. */
52         struct timecounter      *th_counter;
53         int64_t                 th_adjustment;
54         u_int64_t               th_scale;
55         u_int                   th_offset_count;
56         struct bintime          th_offset;
57         struct timeval          th_microtime;
58         struct timespec         th_nanotime;
59         /* Fields not to be copied in tc_windup start with th_generation. */
60         volatile u_int          th_generation;
61         struct timehands        *th_next;
62 };
63
64 static struct timehands th0;
65 static struct timehands th9 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th0};
66 static struct timehands th8 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th9};
67 static struct timehands th7 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th8};
68 static struct timehands th6 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th7};
69 static struct timehands th5 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th6};
70 static struct timehands th4 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th5};
71 static struct timehands th3 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th4};
72 static struct timehands th2 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th3};
73 static struct timehands th1 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th2};
74 static struct timehands th0 = {
75         &dummy_timecounter,
76         0,
77         (uint64_t)-1 / 1000000,
78         0,
79         {1, 0},
80         {0, 0},
81         {0, 0},
82         1,
83         &th1
84 };
85
86 static struct timehands *volatile timehands = &th0;
87 struct timecounter *timecounter = &dummy_timecounter;
88 static struct timecounter *timecounters = &dummy_timecounter;
89
90 time_t time_second = 1;
91 time_t time_uptime = 1;
92
93 static struct bintime boottimebin;
94 struct timeval boottime;
95 static int sysctl_kern_boottime(SYSCTL_HANDLER_ARGS);
96 SYSCTL_PROC(_kern, KERN_BOOTTIME, boottime, CTLTYPE_STRUCT|CTLFLAG_RD,
97     NULL, 0, sysctl_kern_boottime, "S,timeval", "System boottime");
98
99 SYSCTL_NODE(_kern, OID_AUTO, timecounter, CTLFLAG_RW, 0, "");
100
101 static int timestepwarnings;
102 SYSCTL_INT(_kern_timecounter, OID_AUTO, stepwarnings, CTLFLAG_RW,
103     &timestepwarnings, 0, "");
104
105 #define TC_STATS(foo) \
106         static u_int foo; \
107         SYSCTL_UINT(_kern_timecounter, OID_AUTO, foo, CTLFLAG_RD, &foo, 0, "");\
108         struct __hack
109
110 TC_STATS(nbinuptime);    TC_STATS(nnanouptime);    TC_STATS(nmicrouptime);
111 TC_STATS(nbintime);      TC_STATS(nnanotime);      TC_STATS(nmicrotime);
112 TC_STATS(ngetbinuptime); TC_STATS(ngetnanouptime); TC_STATS(ngetmicrouptime);
113 TC_STATS(ngetbintime);   TC_STATS(ngetnanotime);   TC_STATS(ngetmicrotime);
114 TC_STATS(nsetclock);
115
116 #undef TC_STATS
117
118 static void tc_windup(void);
119 static void cpu_tick_calibrate(int);
120
121 static int
122 sysctl_kern_boottime(SYSCTL_HANDLER_ARGS)
123 {
124 #ifdef SCTL_MASK32
125         int tv[2];
126
127         if (req->flags & SCTL_MASK32) {
128                 tv[0] = boottime.tv_sec;
129                 tv[1] = boottime.tv_usec;
130                 return SYSCTL_OUT(req, tv, sizeof(tv));
131         } else
132 #endif
133                 return SYSCTL_OUT(req, &boottime, sizeof(boottime));
134 }
135
136 /*
137  * Return the difference between the timehands' counter value now and what
138  * was when we copied it to the timehands' offset_count.
139  */
140 static __inline u_int
141 tc_delta(struct timehands *th)
142 {
143         struct timecounter *tc;
144
145         tc = th->th_counter;
146         return ((tc->tc_get_timecount(tc) - th->th_offset_count) &
147             tc->tc_counter_mask);
148 }
149
150 /*
151  * Functions for reading the time.  We have to loop until we are sure that
152  * the timehands that we operated on was not updated under our feet.  See
153  * the comment in <sys/time.h> for a description of these 12 functions.
154  */
155
156 void
157 binuptime(struct bintime *bt)
158 {
159         struct timehands *th;
160         u_int gen;
161
162         nbinuptime++;
163         do {
164                 th = timehands;
165                 gen = th->th_generation;
166                 *bt = th->th_offset;
167                 bintime_addx(bt, th->th_scale * tc_delta(th));
168         } while (gen == 0 || gen != th->th_generation);
169 }
170
171 void
172 nanouptime(struct timespec *tsp)
173 {
174         struct bintime bt;
175
176         nnanouptime++;
177         binuptime(&bt);
178         bintime2timespec(&bt, tsp);
179 }
180
181 void
182 microuptime(struct timeval *tvp)
183 {
184         struct bintime bt;
185
186         nmicrouptime++;
187         binuptime(&bt);
188         bintime2timeval(&bt, tvp);
189 }
190
191 void
192 bintime(struct bintime *bt)
193 {
194
195         nbintime++;
196         binuptime(bt);
197         bintime_add(bt, &boottimebin);
198 }
199
200 void
201 nanotime(struct timespec *tsp)
202 {
203         struct bintime bt;
204
205         nnanotime++;
206         bintime(&bt);
207         bintime2timespec(&bt, tsp);
208 }
209
210 void
211 microtime(struct timeval *tvp)
212 {
213         struct bintime bt;
214
215         nmicrotime++;
216         bintime(&bt);
217         bintime2timeval(&bt, tvp);
218 }
219
220 void
221 getbinuptime(struct bintime *bt)
222 {
223         struct timehands *th;
224         u_int gen;
225
226         ngetbinuptime++;
227         do {
228                 th = timehands;
229                 gen = th->th_generation;
230                 *bt = th->th_offset;
231         } while (gen == 0 || gen != th->th_generation);
232 }
233
234 void
235 getnanouptime(struct timespec *tsp)
236 {
237         struct timehands *th;
238         u_int gen;
239
240         ngetnanouptime++;
241         do {
242                 th = timehands;
243                 gen = th->th_generation;
244                 bintime2timespec(&th->th_offset, tsp);
245         } while (gen == 0 || gen != th->th_generation);
246 }
247
248 void
249 getmicrouptime(struct timeval *tvp)
250 {
251         struct timehands *th;
252         u_int gen;
253
254         ngetmicrouptime++;
255         do {
256                 th = timehands;
257                 gen = th->th_generation;
258                 bintime2timeval(&th->th_offset, tvp);
259         } while (gen == 0 || gen != th->th_generation);
260 }
261
262 void
263 getbintime(struct bintime *bt)
264 {
265         struct timehands *th;
266         u_int gen;
267
268         ngetbintime++;
269         do {
270                 th = timehands;
271                 gen = th->th_generation;
272                 *bt = th->th_offset;
273         } while (gen == 0 || gen != th->th_generation);
274         bintime_add(bt, &boottimebin);
275 }
276
277 void
278 getnanotime(struct timespec *tsp)
279 {
280         struct timehands *th;
281         u_int gen;
282
283         ngetnanotime++;
284         do {
285                 th = timehands;
286                 gen = th->th_generation;
287                 *tsp = th->th_nanotime;
288         } while (gen == 0 || gen != th->th_generation);
289 }
290
291 void
292 getmicrotime(struct timeval *tvp)
293 {
294         struct timehands *th;
295         u_int gen;
296
297         ngetmicrotime++;
298         do {
299                 th = timehands;
300                 gen = th->th_generation;
301                 *tvp = th->th_microtime;
302         } while (gen == 0 || gen != th->th_generation);
303 }
304
305 /*
306  * Initialize a new timecounter and possibly use it.
307  */
308 void
309 tc_init(struct timecounter *tc)
310 {
311         u_int u;
312
313         u = tc->tc_frequency / tc->tc_counter_mask;
314         /* XXX: We need some margin here, 10% is a guess */
315         u *= 11;
316         u /= 10;
317         if (u > hz && tc->tc_quality >= 0) {
318                 tc->tc_quality = -2000;
319                 if (bootverbose) {
320                         printf("Timecounter \"%s\" frequency %ju Hz",
321                             tc->tc_name, (uintmax_t)tc->tc_frequency);
322                         printf(" -- Insufficient hz, needs at least %u\n", u);
323                 }
324         } else if (tc->tc_quality >= 0 || bootverbose) {
325                 printf("Timecounter \"%s\" frequency %ju Hz quality %d\n",
326                     tc->tc_name, (uintmax_t)tc->tc_frequency,
327                     tc->tc_quality);
328         }
329
330         tc->tc_next = timecounters;
331         timecounters = tc;
332         /*
333          * Never automatically use a timecounter with negative quality.
334          * Even though we run on the dummy counter, switching here may be
335          * worse since this timecounter may not be monotonous.
336          */
337         if (tc->tc_quality < 0)
338                 return;
339         if (tc->tc_quality < timecounter->tc_quality)
340                 return;
341         if (tc->tc_quality == timecounter->tc_quality &&
342             tc->tc_frequency < timecounter->tc_frequency)
343                 return;
344         (void)tc->tc_get_timecount(tc);
345         (void)tc->tc_get_timecount(tc);
346         timecounter = tc;
347 }
348
349 /* Report the frequency of the current timecounter. */
350 u_int64_t
351 tc_getfrequency(void)
352 {
353
354         return (timehands->th_counter->tc_frequency);
355 }
356
357 /*
358  * Step our concept of UTC.  This is done by modifying our estimate of
359  * when we booted.
360  * XXX: not locked.
361  */
362 void
363 tc_setclock(struct timespec *ts)
364 {
365         struct timespec tbef, taft;
366         struct bintime bt, bt2;
367
368         cpu_tick_calibrate(1);
369         nsetclock++;
370         nanotime(&tbef);
371         timespec2bintime(ts, &bt);
372         binuptime(&bt2);
373         bintime_sub(&bt, &bt2);
374         bintime_add(&bt2, &boottimebin);
375         boottimebin = bt;
376         bintime2timeval(&bt, &boottime);
377
378         /* XXX fiddle all the little crinkly bits around the fiords... */
379         tc_windup();
380         nanotime(&taft);
381         if (timestepwarnings) {
382                 log(LOG_INFO,
383                     "Time stepped from %jd.%09ld to %jd.%09ld (%jd.%09ld)\n",
384                     (intmax_t)tbef.tv_sec, tbef.tv_nsec,
385                     (intmax_t)taft.tv_sec, taft.tv_nsec,
386                     (intmax_t)ts->tv_sec, ts->tv_nsec);
387         }
388         cpu_tick_calibrate(1);
389 }
390
391 /*
392  * Initialize the next struct timehands in the ring and make
393  * it the active timehands.  Along the way we might switch to a different
394  * timecounter and/or do seconds processing in NTP.  Slightly magic.
395  */
396 static void
397 tc_windup(void)
398 {
399         struct bintime bt;
400         struct timehands *th, *tho;
401         u_int64_t scale;
402         u_int delta, ncount, ogen;
403         int i;
404         time_t t;
405
406         /*
407          * Make the next timehands a copy of the current one, but do not
408          * overwrite the generation or next pointer.  While we update
409          * the contents, the generation must be zero.
410          */
411         tho = timehands;
412         th = tho->th_next;
413         ogen = th->th_generation;
414         th->th_generation = 0;
415         bcopy(tho, th, offsetof(struct timehands, th_generation));
416
417         /*
418          * Capture a timecounter delta on the current timecounter and if
419          * changing timecounters, a counter value from the new timecounter.
420          * Update the offset fields accordingly.
421          */
422         delta = tc_delta(th);
423         if (th->th_counter != timecounter)
424                 ncount = timecounter->tc_get_timecount(timecounter);
425         else
426                 ncount = 0;
427         th->th_offset_count += delta;
428         th->th_offset_count &= th->th_counter->tc_counter_mask;
429         bintime_addx(&th->th_offset, th->th_scale * delta);
430
431         /*
432          * Hardware latching timecounters may not generate interrupts on
433          * PPS events, so instead we poll them.  There is a finite risk that
434          * the hardware might capture a count which is later than the one we
435          * got above, and therefore possibly in the next NTP second which might
436          * have a different rate than the current NTP second.  It doesn't
437          * matter in practice.
438          */
439         if (tho->th_counter->tc_poll_pps)
440                 tho->th_counter->tc_poll_pps(tho->th_counter);
441
442         /*
443          * Deal with NTP second processing.  The for loop normally
444          * iterates at most once, but in extreme situations it might
445          * keep NTP sane if timeouts are not run for several seconds.
446          * At boot, the time step can be large when the TOD hardware
447          * has been read, so on really large steps, we call
448          * ntp_update_second only twice.  We need to call it twice in
449          * case we missed a leap second.
450          */
451         bt = th->th_offset;
452         bintime_add(&bt, &boottimebin);
453         i = bt.sec - tho->th_microtime.tv_sec;
454         if (i > LARGE_STEP)
455                 i = 2;
456         for (; i > 0; i--) {
457                 t = bt.sec;
458                 ntp_update_second(&th->th_adjustment, &bt.sec);
459                 if (bt.sec != t)
460                         boottimebin.sec += bt.sec - t;
461         }
462         /* Update the UTC timestamps used by the get*() functions. */
463         /* XXX shouldn't do this here.  Should force non-`get' versions. */
464         bintime2timeval(&bt, &th->th_microtime);
465         bintime2timespec(&bt, &th->th_nanotime);
466
467         /* Now is a good time to change timecounters. */
468         if (th->th_counter != timecounter) {
469                 th->th_counter = timecounter;
470                 th->th_offset_count = ncount;
471         }
472
473         /*-
474          * Recalculate the scaling factor.  We want the number of 1/2^64
475          * fractions of a second per period of the hardware counter, taking
476          * into account the th_adjustment factor which the NTP PLL/adjtime(2)
477          * processing provides us with.
478          *
479          * The th_adjustment is nanoseconds per second with 32 bit binary
480          * fraction and we want 64 bit binary fraction of second:
481          *
482          *       x = a * 2^32 / 10^9 = a * 4.294967296
483          *
484          * The range of th_adjustment is +/- 5000PPM so inside a 64bit int
485          * we can only multiply by about 850 without overflowing, that
486          * leaves no suitably precise fractions for multiply before divide.
487          *
488          * Divide before multiply with a fraction of 2199/512 results in a
489          * systematic undercompensation of 10PPM of th_adjustment.  On a
490          * 5000PPM adjustment this is a 0.05PPM error.  This is acceptable.
491          *
492          * We happily sacrifice the lowest of the 64 bits of our result
493          * to the goddess of code clarity.
494          *
495          */
496         scale = (u_int64_t)1 << 63;
497         scale += (th->th_adjustment / 1024) * 2199;
498         scale /= th->th_counter->tc_frequency;
499         th->th_scale = scale * 2;
500
501         /*
502          * Now that the struct timehands is again consistent, set the new
503          * generation number, making sure to not make it zero.
504          */
505         if (++ogen == 0)
506                 ogen = 1;
507         th->th_generation = ogen;
508
509         /* Go live with the new struct timehands. */
510         time_second = th->th_microtime.tv_sec;
511         time_uptime = th->th_offset.sec;
512         timehands = th;
513 }
514
515 /* Report or change the active timecounter hardware. */
516 static int
517 sysctl_kern_timecounter_hardware(SYSCTL_HANDLER_ARGS)
518 {
519         char newname[32];
520         struct timecounter *newtc, *tc;
521         int error;
522
523         tc = timecounter;
524         strlcpy(newname, tc->tc_name, sizeof(newname));
525
526         error = sysctl_handle_string(oidp, &newname[0], sizeof(newname), req);
527         if (error != 0 || req->newptr == NULL ||
528             strcmp(newname, tc->tc_name) == 0)
529                 return (error);
530         for (newtc = timecounters; newtc != NULL; newtc = newtc->tc_next) {
531                 if (strcmp(newname, newtc->tc_name) != 0)
532                         continue;
533
534                 /* Warm up new timecounter. */
535                 (void)newtc->tc_get_timecount(newtc);
536                 (void)newtc->tc_get_timecount(newtc);
537
538                 timecounter = newtc;
539                 return (0);
540         }
541         return (EINVAL);
542 }
543
544 SYSCTL_PROC(_kern_timecounter, OID_AUTO, hardware, CTLTYPE_STRING | CTLFLAG_RW,
545     0, 0, sysctl_kern_timecounter_hardware, "A", "");
546
547
548 /* Report or change the active timecounter hardware. */
549 static int
550 sysctl_kern_timecounter_choice(SYSCTL_HANDLER_ARGS)
551 {
552         char buf[32], *spc;
553         struct timecounter *tc;
554         int error;
555
556         spc = "";
557         error = 0;
558         for (tc = timecounters; error == 0 && tc != NULL; tc = tc->tc_next) {
559                 sprintf(buf, "%s%s(%d)",
560                     spc, tc->tc_name, tc->tc_quality);
561                 error = SYSCTL_OUT(req, buf, strlen(buf));
562                 spc = " ";
563         }
564         return (error);
565 }
566
567 SYSCTL_PROC(_kern_timecounter, OID_AUTO, choice, CTLTYPE_STRING | CTLFLAG_RD,
568     0, 0, sysctl_kern_timecounter_choice, "A", "");
569
570 /*
571  * RFC 2783 PPS-API implementation.
572  */
573
574 int
575 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
576 {
577         pps_params_t *app;
578         struct pps_fetch_args *fapi;
579 #ifdef PPS_SYNC
580         struct pps_kcbind_args *kapi;
581 #endif
582
583         KASSERT(pps != NULL, ("NULL pps pointer in pps_ioctl"));
584         switch (cmd) {
585         case PPS_IOC_CREATE:
586                 return (0);
587         case PPS_IOC_DESTROY:
588                 return (0);
589         case PPS_IOC_SETPARAMS:
590                 app = (pps_params_t *)data;
591                 if (app->mode & ~pps->ppscap)
592                         return (EINVAL);
593                 pps->ppsparam = *app;
594                 return (0);
595         case PPS_IOC_GETPARAMS:
596                 app = (pps_params_t *)data;
597                 *app = pps->ppsparam;
598                 app->api_version = PPS_API_VERS_1;
599                 return (0);
600         case PPS_IOC_GETCAP:
601                 *(int*)data = pps->ppscap;
602                 return (0);
603         case PPS_IOC_FETCH:
604                 fapi = (struct pps_fetch_args *)data;
605                 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
606                         return (EINVAL);
607                 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec)
608                         return (EOPNOTSUPP);
609                 pps->ppsinfo.current_mode = pps->ppsparam.mode;
610                 fapi->pps_info_buf = pps->ppsinfo;
611                 return (0);
612         case PPS_IOC_KCBIND:
613 #ifdef PPS_SYNC
614                 kapi = (struct pps_kcbind_args *)data;
615                 /* XXX Only root should be able to do this */
616                 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
617                         return (EINVAL);
618                 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
619                         return (EINVAL);
620                 if (kapi->edge & ~pps->ppscap)
621                         return (EINVAL);
622                 pps->kcmode = kapi->edge;
623                 return (0);
624 #else
625                 return (EOPNOTSUPP);
626 #endif
627         default:
628                 return (ENOIOCTL);
629         }
630 }
631
632 void
633 pps_init(struct pps_state *pps)
634 {
635         pps->ppscap |= PPS_TSFMT_TSPEC;
636         if (pps->ppscap & PPS_CAPTUREASSERT)
637                 pps->ppscap |= PPS_OFFSETASSERT;
638         if (pps->ppscap & PPS_CAPTURECLEAR)
639                 pps->ppscap |= PPS_OFFSETCLEAR;
640 }
641
642 void
643 pps_capture(struct pps_state *pps)
644 {
645         struct timehands *th;
646
647         KASSERT(pps != NULL, ("NULL pps pointer in pps_capture"));
648         th = timehands;
649         pps->capgen = th->th_generation;
650         pps->capth = th;
651         pps->capcount = th->th_counter->tc_get_timecount(th->th_counter);
652         if (pps->capgen != th->th_generation)
653                 pps->capgen = 0;
654 }
655
656 void
657 pps_event(struct pps_state *pps, int event)
658 {
659         struct bintime bt;
660         struct timespec ts, *tsp, *osp;
661         u_int tcount, *pcount;
662         int foff, fhard;
663         pps_seq_t *pseq;
664
665         KASSERT(pps != NULL, ("NULL pps pointer in pps_event"));
666         /* If the timecounter was wound up underneath us, bail out. */
667         if (pps->capgen == 0 || pps->capgen != pps->capth->th_generation)
668                 return;
669
670         /* Things would be easier with arrays. */
671         if (event == PPS_CAPTUREASSERT) {
672                 tsp = &pps->ppsinfo.assert_timestamp;
673                 osp = &pps->ppsparam.assert_offset;
674                 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
675                 fhard = pps->kcmode & PPS_CAPTUREASSERT;
676                 pcount = &pps->ppscount[0];
677                 pseq = &pps->ppsinfo.assert_sequence;
678         } else {
679                 tsp = &pps->ppsinfo.clear_timestamp;
680                 osp = &pps->ppsparam.clear_offset;
681                 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
682                 fhard = pps->kcmode & PPS_CAPTURECLEAR;
683                 pcount = &pps->ppscount[1];
684                 pseq = &pps->ppsinfo.clear_sequence;
685         }
686
687         /*
688          * If the timecounter changed, we cannot compare the count values, so
689          * we have to drop the rest of the PPS-stuff until the next event.
690          */
691         if (pps->ppstc != pps->capth->th_counter) {
692                 pps->ppstc = pps->capth->th_counter;
693                 *pcount = pps->capcount;
694                 pps->ppscount[2] = pps->capcount;
695                 return;
696         }
697
698         /* Convert the count to a timespec. */
699         tcount = pps->capcount - pps->capth->th_offset_count;
700         tcount &= pps->capth->th_counter->tc_counter_mask;
701         bt = pps->capth->th_offset;
702         bintime_addx(&bt, pps->capth->th_scale * tcount);
703         bintime_add(&bt, &boottimebin);
704         bintime2timespec(&bt, &ts);
705
706         /* If the timecounter was wound up underneath us, bail out. */
707         if (pps->capgen != pps->capth->th_generation)
708                 return;
709
710         *pcount = pps->capcount;
711         (*pseq)++;
712         *tsp = ts;
713
714         if (foff) {
715                 timespecadd(tsp, osp);
716                 if (tsp->tv_nsec < 0) {
717                         tsp->tv_nsec += 1000000000;
718                         tsp->tv_sec -= 1;
719                 }
720         }
721 #ifdef PPS_SYNC
722         if (fhard) {
723                 u_int64_t scale;
724
725                 /*
726                  * Feed the NTP PLL/FLL.
727                  * The FLL wants to know how many (hardware) nanoseconds
728                  * elapsed since the previous event.
729                  */
730                 tcount = pps->capcount - pps->ppscount[2];
731                 pps->ppscount[2] = pps->capcount;
732                 tcount &= pps->capth->th_counter->tc_counter_mask;
733                 scale = (u_int64_t)1 << 63;
734                 scale /= pps->capth->th_counter->tc_frequency;
735                 scale *= 2;
736                 bt.sec = 0;
737                 bt.frac = 0;
738                 bintime_addx(&bt, scale * tcount);
739                 bintime2timespec(&bt, &ts);
740                 hardpps(tsp, ts.tv_nsec + 1000000000 * ts.tv_sec);
741         }
742 #endif
743 }
744
745 /*
746  * Timecounters need to be updated every so often to prevent the hardware
747  * counter from overflowing.  Updating also recalculates the cached values
748  * used by the get*() family of functions, so their precision depends on
749  * the update frequency.
750  */
751
752 static int tc_tick;
753 SYSCTL_INT(_kern_timecounter, OID_AUTO, tick, CTLFLAG_RD, &tc_tick, 0, "");
754
755 void
756 tc_ticktock(void)
757 {
758         static int count;
759         static time_t last_calib;
760
761         if (++count < tc_tick)
762                 return;
763         count = 0;
764         tc_windup();
765         if (time_uptime != last_calib && !(time_uptime & 0xf)) {
766                 cpu_tick_calibrate(0);
767                 last_calib = time_uptime;
768         }
769 }
770
771 static void
772 inittimecounter(void *dummy)
773 {
774         u_int p;
775
776         /*
777          * Set the initial timeout to
778          * max(1, <approx. number of hardclock ticks in a millisecond>).
779          * People should probably not use the sysctl to set the timeout
780          * to smaller than its inital value, since that value is the
781          * smallest reasonable one.  If they want better timestamps they
782          * should use the non-"get"* functions.
783          */
784         if (hz > 1000)
785                 tc_tick = (hz + 500) / 1000;
786         else
787                 tc_tick = 1;
788         p = (tc_tick * 1000000) / hz;
789         printf("Timecounters tick every %d.%03u msec\n", p / 1000, p % 1000);
790
791         /* warm up new timecounter (again) and get rolling. */
792         (void)timecounter->tc_get_timecount(timecounter);
793         (void)timecounter->tc_get_timecount(timecounter);
794 }
795
796 SYSINIT(timecounter, SI_SUB_CLOCKS, SI_ORDER_SECOND, inittimecounter, NULL)
797
798 /* Cpu tick handling -------------------------------------------------*/
799
800 static int cpu_tick_variable;
801 static uint64_t cpu_tick_frequency;
802
803 static
804 uint64_t
805 tc_cpu_ticks(void)
806 {
807         static uint64_t base;
808         static unsigned last;
809         unsigned u;
810         struct timecounter *tc;
811
812         tc = timehands->th_counter;
813         u = tc->tc_get_timecount(tc) & tc->tc_counter_mask;
814         if (u < last)
815                 base += (uint64_t)tc->tc_counter_mask + 1;
816         last = u;
817         return (u + base);
818 }
819
820 /*
821  * This function gets called ever 16 seconds on only one designated
822  * CPU in the system from hardclock() via tc_ticktock().
823  *
824  * Whenever the real time clock is stepped we get called with reset=1
825  * to make sure we handle suspend/resume and similar events correctly.
826  */
827
828 static void
829 cpu_tick_calibrate(int reset)
830 {
831         static uint64_t c_last;
832         uint64_t c_this, c_delta;
833         static struct bintime  t_last;
834         struct bintime t_this, t_delta;
835         uint32_t divi;
836
837         if (reset) {
838                 /* The clock was stepped, abort & reset */
839                 t_last.sec = 0;
840                 return;
841         }
842
843         /* we don't calibrate fixed rate cputicks */
844         if (!cpu_tick_variable)
845                 return;
846
847         getbinuptime(&t_this);
848         c_this = cpu_ticks();
849         if (t_last.sec != 0) {
850                 c_delta = c_this - c_last;
851                 t_delta = t_this;
852                 bintime_sub(&t_delta, &t_last);
853                 if (bootverbose) {
854                         printf("%ju.%016jx ",
855                             (uintmax_t)t_delta.sec, (uintmax_t)t_delta.frac);
856                 }
857                 /*
858                  * Validate that 16 +/- 1/256 seconds passed. 
859                  * After division by 16 this gives us a precision of
860                  * roughly 250PPM which is sufficient
861                  */
862                 if (t_delta.sec > 16 || (
863                     t_delta.sec == 16 && t_delta.frac >= (0x01LL << 56))) {
864                         /* too long */
865                         if (bootverbose)
866                                 printf("\ttoo long\n");
867                 } else if (t_delta.sec < 15 ||
868                     (t_delta.sec == 15 && t_delta.frac <= (0xffLL << 56))) {
869                         /* too short */
870                         if (bootverbose)
871                                 printf("\ttoo short\n");
872                 } else {
873                         /* just right */
874                         /*
875                          * Headroom:
876                          *      2^(64-20) / 16[s] =
877                          *      2^(44) / 16[s] =
878                          *      17.592.186.044.416 / 16 =
879                          *      1.099.511.627.776 [Hz]
880                          */
881                         divi = t_delta.sec << 20;
882                         divi |= t_delta.frac >> (64 - 20);
883                         c_delta <<= 20;
884                         if (bootverbose)
885                                 printf(" %ju / %ju",
886                                     (uintmax_t)c_delta, (uintmax_t)divi);
887                         c_delta /= divi;
888                         if (bootverbose)
889                                 printf(" = %ju", c_delta);
890                         if (c_delta  > cpu_tick_frequency) {
891                                 if (bootverbose)
892                                         printf("\thigher\n");
893                                 cpu_tick_frequency = c_delta;
894                         } else {
895                                 if (bootverbose)
896                                         printf("\tlower\n");
897                         }
898                 }
899         }
900         c_last = c_this;
901         t_last = t_this;
902 }
903
904 void
905 set_cputicker(cpu_tick_f *func, uint64_t freq, unsigned var)
906 {
907
908         if (func == NULL) {
909                 cpu_ticks = tc_cpu_ticks;
910         } else {
911                 cpu_tick_frequency = freq;
912                 cpu_tick_variable = var;
913                 cpu_ticks = func;
914         }
915 }
916
917 uint64_t
918 cpu_tickrate(void)
919 {
920
921         if (cpu_ticks == tc_cpu_ticks) 
922                 return (tc_getfrequency());
923         return (cpu_tick_frequency);
924 }
925
926 /*
927  * We need to be slightly careful converting cputicks to microseconds.
928  * There is plenty of margin in 64 bits of microseconds (half a million
929  * years) and in 64 bits at 4 GHz (146 years), but if we do a multiply
930  * before divide conversion (to retain precision) we find that the
931  * margin shrinks to 1.5 hours (one millionth of 146y).
932  * With a three prong approach we never loose significant bits, no
933  * matter what the cputick rate and length of timeinterval is.
934  */
935
936 uint64_t
937 cputick2usec(uint64_t tick)
938 {
939
940         if (tick > 18446744073709551LL)         /* floor(2^64 / 1000) */
941                 return (tick / (cpu_tickrate() / 1000000LL));
942         else if (tick > 18446744073709LL)       /* floor(2^64 / 1000000) */
943                 return ((tick * 1000LL) / (cpu_tickrate() / 1000LL));
944         else
945                 return ((tick * 1000000LL) / cpu_tickrate());
946 }
947
948 cpu_tick_f      *cpu_ticks = tc_cpu_ticks;